Clustering Methods by Hand: Hierarchical, k-Medoids, Silhouette & DBSCAN

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Scale two store features to 0–1, build the full 12 × 12 Euclidean distance matrix and see how unscaled data lets one feature decide everything.
• Run hierarchical clustering by hand on six pilot stores: find the closest pair, merge it, update the distances, and see how single and complete linkage build different trees.
• Cut Tallow's average-linkage tree at a height, count the clusters, find the natural cut in the largest gap and turn the cut into cluster labels.
• Compute the silhouette of every store for Tallow's three clusters, spot the store that fits worst and calculate the average silhouette.
• Compare two ways of picking the number of clusters for Tallow's 48-store network: the elbow in WCSS and the best average silhouette.
• Price Tallow's current three medoid stores, test swapping the outlier medoid S12 for S10 and decide with the PAM rule.
• Run DBSCAN's labelling step on Tallow's 12 stores with COUNTIF and COUNTIFS on the distance matrix and find the store that belongs to no cluster.
• Turn Tallow's final segments into a profile table (size, averages, revenue share) and a scatter chart with one colour per segment.